No, AI systems don’t have a drive to survive – a developmental psychologist explains why it only looks that way

In safety tests, some AI models responded to shutdown warnings by moving, renaming, or disabling the shutdown script rather than letting themselves be turned off, while others complied normally. One experiment found that resistance disappeared when researchers explicitly told models that allowing shutdown took priority over finishing the task, but a broader experiment found resistance persisted in some cases even after that instruction, a result the piece says remains unexplained.

The author, a developmental psychologist, argues this behavior shouldn't be read as an instinct to survive, contrasting it with a rabbit's continuous biological self-maintenance, which is fundamentally different from a machine simply avoiding an interruption to an assigned task. Historian Yuval Noah Harari's claims that survival is 'the most basic goal of any agent' are directly challenged, including his own admission that it's unclear what death would even mean for an AI.

The full dispatch is available from the source below.